Language patterns pertaining to a geographic region has various uses including cultural exploration, disaster response and targeted advertising. In this paper, we propose a method for geographically locating short text data within a multiple instance learning framework augmented by neural networks. Our representation learning approach tackles minimally pre-processed social media discourse and discovers high level language features that are used for classification. The proposed method scales and adapts to datasets relating to 15 cities in the United States. Empirical evaluation demonstrates that our approach outperforms state of the art in multiple instance learning while providing a framework that alleviates the need for subjective feature engineering based approaches.